📈 economics

Dynamic Cheap Talk without Feedback

This paper demonstrates that in a dynamic sender-receiver game with a Markovian state and no feedback, the sender can partially restore commitment power to achieve equilibrium payoffs equivalent to those of a persuasion model with partial commitment, including the Bayesian persuasion payoff when the sender's utility is state-independent.

Original authors: Atulya Jain

Published 2026-04-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Atulya Jain

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a game of "Telephone" played between two people: an Expert (the Sender) who knows the secret truth about the world, and a Decision-Maker (the Receiver) who has to make choices based on what the Expert tells them.

Usually, this game is tricky. The Expert might lie to get a better result for themselves, and the Decision-Maker knows this, so they don't fully trust the Expert. This limits how much useful information can be shared.

In many real-life situations, like a financial advisor talking to an investor or a doctor talking to a patient, the Expert gives advice but never sees what happens next. They don't know if the advice was followed, nor do they see the final result. In game theory, this is called "no feedback." Usually, without feedback, it's very hard to keep a liar in line because there's no way to punish them later.

The Big Surprise
This paper argues that even without feedback, playing this game over and over again (dynamically) can actually help the Expert tell the truth more often than they could in a single, one-time conversation.

Here is the simple breakdown of how it works, using some everyday analogies:

1. The "Quota" System (The Main Trick)

The paper suggests a clever way to keep the Expert honest without needing to see the results. Imagine the game is played in long blocks of time (like a school semester).

The Rule: Before the block starts, the two players agree on a "quota" for how many times each type of message should be sent. For example, in a 100-day block, the Expert must say "Good News" exactly 40 times and "Bad News" exactly 60 times.
The Monitoring: The Decision-Maker keeps a tally. As long as the Expert sticks to the quota, the Decision-Maker trusts the message and acts on it.
The Punishment: If the Expert tries to cheat by saying "Good News" too many times early on, the Decision-Maker notices the quota is filling up too fast. They then stop listening to the Expert's current message and switch to a "safe" message that hasn't been used up yet.

Why this works: The Expert knows that if they lie too much now, they will run out of "Good News" cards later in the block. Since they need to maintain the overall balance (the quota) to keep the Decision-Maker trusting them, they are forced to tell the truth more often. It's like a teacher who knows a student must write exactly 5 essays on "History" and 5 on "Science" to pass; the student can't just write 10 on History to get an easy A.

2. The "Magic" of No Feedback

Usually, you need to see the outcome to know if someone is cheating. But here, the Decision-Maker doesn't need to see the outcome. They only need to count the messages.

Because the Expert doesn't know what the Decision-Maker is doing, they can't adjust their strategy based on the Decision-Maker's reaction. However, the Decision-Maker can see the pattern of messages. If the Expert tries to change the pattern (the distribution of messages), the Decision-Maker catches them statistically. This forces the Expert to stick to a specific "script" of messages, which partially restores their ability to commit to telling the truth.

3. The Special Case: When the Expert Doesn't Care About the State

The paper finds a special scenario where this dynamic game is incredibly powerful. Imagine the Expert's goal is completely independent of the actual state of the world.

Example: A doctor wants to prescribe a specific drug because it's profitable for the hospital, regardless of whether the patient actually has the disease.

In a one-time conversation, the doctor might lie about the disease to get the patient to take the drug. But in this dynamic "no feedback" game, the paper shows the doctor can achieve the absolute best possible outcome (the same result as if they could magically force the patient to believe them).

Even though the doctor can't see if the patient took the medicine, the "quota" system forces them to reveal the truth about the disease distribution perfectly. The dynamic interaction bridges the gap between "cheap talk" (lying is easy) and "full commitment" (lying is impossible).

4. The Limits (When it doesn't work perfectly)

The paper also admits this isn't a magic wand for every situation.

The "Copycat" Problem: In some complex scenarios, the Expert might try to cheat in a way that keeps the total number of messages the same but changes the timing or sequence (e.g., saying "Good" then "Bad" instead of "Bad" then "Good").
The paper shows that sometimes the Decision-Maker can catch these timing tricks, and sometimes they can't. If the state of the world changes randomly and independently every day (like flipping a coin), the "quota" system works perfectly. But if the state has a memory (like the weather, where today affects tomorrow), the system is slightly less perfect, though still better than a one-time conversation.

Summary

The Problem: Experts often lie because they can't be punished for it if they don't see the results.
The Solution: Play the game in long blocks with pre-set "message quotas."
The Result: The Expert is forced to tell the truth to keep their message count balanced. This allows them to achieve much better outcomes than in a single conversation, and in some cases, they can achieve the same perfect results as if they had a magical ability to commit to the truth.
The Catch: It works best when the world changes randomly; if the world has complex patterns, the system is slightly less perfect but still an improvement.

The paper essentially proves that repetition and statistical monitoring can substitute for direct observation and punishment, allowing for better communication even when the expert is flying blind.

1. Problem Statement

The paper investigates strategic information transmission in a dynamic sender-receiver game where:

Players: An informed Sender (expert) and an uninformed Receiver (decision-maker).
State Dynamics: The state of the world $\omega_n$ evolves over time according to a Markov chain (specifically, a pseudo-renewal chain).
Information Structure: The Sender observes the current state privately and sends a message. The Receiver chooses an action based on the message.
Crucial Constraint (No Feedback): The Sender does not observe the Receiver's action or the realized payoff. The Receiver observes the history of messages but not the state.
Objective: To characterize the set of uniform equilibrium payoffs (robust limits of approximate equilibria for sufficiently patient players) and determine how much the "commitment gap" (the difference between cheap talk and Bayesian persuasion) can be closed without feedback.

2. Methodology and Framework

A. The Model

Game Structure: Infinite horizon, discrete time.
Markov Chain: The state follows an irreducible, aperiodic, pseudo-renewal Markov chain. A chain is pseudo-renewal if the transition probability to a destination state depends only on the destination, not the current state (e.g., i.i.d. processes or specific shock-driven processes).
Solution Concept: Uniform Equilibrium. Since players do not observe stage payoffs, standard Perfect Bayesian Equilibrium (PBE) is sensitive to perturbations. The author focuses on uniform equilibria, defined as the limit of $\epsilon$ -equilibria sustained by a single strategy profile for all sufficiently high discount factors $\delta$ .

B. The Static Benchmark: Partial Commitment

The author introduces a static persuasion model with partial commitment as the primary benchmark.

Constraint: The Sender can choose any signaling policy, provided it preserves a pre-specified marginal distribution over messages ( $\lambda$ ).
Rationale: In the dynamic game, the Receiver can monitor the empirical frequency of messages. Deviations that alter the marginal distribution $\lambda$ are detectable and punishable.
Equilibrium Conditions:
1. Receiver: Best responds to posterior beliefs induced by messages.
2. Sender: Cannot profitably deviate to any other policy that preserves the marginal distribution $\lambda$ .
Notation: Let $E(\pi, \lambda)$ be the set of equilibrium payoffs for prior $\pi$ and marginal $\lambda$ .

C. The Dynamic Construction: Block Strategies and Quotas

To achieve these static payoffs dynamically, the author constructs a strategy profile using block strategies:

Blocks: Time is divided into blocks of length $N$ .
Quota Mechanism: The Receiver enforces a quota for each message $m$ $m$ within a block, specifically $N \cdot \lambda(m)$ $N \cdot λ (m)$ .
- If the Sender sends a message $m$ and its quota is not yet filled, the Receiver plays the action corresponding to the equilibrium response for $m$ .
- If the quota for $m$ is filled, the Receiver ignores the Sender's message and plays the action for a different message $m'$ whose quota is not yet filled.
Sender's Strategy: The Sender uses a recursive strategy to induce the target posterior beliefs $\{p_m\}$ while ensuring the empirical frequency of messages matches $\lambda$ within the block.
Pseudo-renewal Property: This specific Markov structure allows the Sender to construct a strategy where the belief evolution is consistent with the target posteriors, even though the Sender does not observe the Receiver's actions.

3. Key Contributions and Results

A. Main Theorem (Theorem 1): Attainability

Result: Any equilibrium payoff of the static persuasion model with partial commitment (evaluated at the invariant distribution $\mu$ ), and any convex combination of such payoffs across different message distributions, can be achieved as a uniform equilibrium payoff in the dynamic game.

Mathematical Form: $Co_\lambda(E(\mu, \lambda)) \subseteq U$ , where $U$ is the set of uniform equilibrium payoffs.
Implication: Dynamic interaction without feedback partially restores commitment. The Sender can achieve payoffs strictly higher than in static cheap talk, approaching the Bayesian persuasion payoff.

B. Special Case: State-Independent Preferences

Result: If the Sender's payoff is state-independent (i.e., depends only on the Receiver's action, not the state), the Sender achieves the Bayesian persuasion payoff (the theoretical upper bound).

Significance: This contrasts with previous literature (e.g., Kuvalekar et al., 2022) where repetition without feedback failed to restore full commitment even in i.i.d. settings. Here, the Markovian structure and the specific "partial commitment" constraint allow the Sender to fully bridge the gap.

C. Necessity and Limits (Proposition 1 & Example 2)

Result: The converse of Theorem 1 does not hold in general. The set of uniform equilibrium payoffs is not necessarily contained within the convex hull of partial commitment payoffs.

Reasoning: In the dynamic game, the Receiver can perform richer statistical tests (e.g., checking the frequency of pairs of messages or temporal correlations) that are not detectable in the static model.
Example 2: The author constructs a scenario where a dynamic equilibrium yields a payoff of $1/3$ , while every static partial-commitment equilibrium yields at least $1/2$ . This is because the dynamic strategy relies on specific temporal correlations (e.g., message $m$ in period $t$ must be followed by $m$ in $t+1$ ) that are detectable deviations in the dynamic setting but invisible in the static marginal constraint.
Necessary Condition (Proposition 1): Any uniform equilibrium payoff must be robust to copula deviations. These are deviations where the Sender applies the equilibrium strategy to a fictitious state process that is statistically indistinguishable from the true process (sharing the same marginal distribution but potentially different joint structure).

D. Exact Characterization for i.i.d. States (Proposition 2)

Result: If the state process is i.i.d. (a special case of pseudo-renewal), the set of uniform equilibrium payoffs coincides exactly with the convex hull of the partial commitment payoffs:
$U = Co_\lambda(E(\mu, \lambda))$

Significance: In the i.i.d. case, temporal correlations vanish, making the static partial-commitment benchmark a complete micro-foundation for the dynamic game.

4. Significance and Contribution to Literature

Restoring Commitment without Feedback: The paper demonstrates that dynamic interaction alone can partially restore commitment even when the Sender receives no feedback. This challenges the intuition that feedback is strictly necessary for reputation or punishment mechanisms.
Micro-foundation for Partial Commitment: It provides a game-theoretic justification for the "partial commitment" models (like Lin and Liu, 2024) often used in static analysis. The restriction on marginal message distributions arises endogenously from the Receiver's ability to monitor frequencies.
Distinction from Existing Literature:
- Contrasts with Kuvalekar, Lipnowski, and Ramos (2022): They show that in i.i.d. cheap talk without feedback, the Bayesian persuasion payoff is unattainable. This paper shows that with Markovian states and uniform equilibrium concepts, the upper bound is attainable for state-independent payoffs.
- Extends Renault, Solan, and Vieille (2013): While they studied similar mechanisms, this paper specifically addresses the sender-receiver communication problem and the specific constraints of "cheap talk" (no money burning).
Methodological Innovation: The use of quota mechanisms combined with pseudo-renewal Markov chains to construct recursive belief systems that satisfy the Sender's incentive constraints without feedback is a novel technical contribution.

5. Conclusion

The paper establishes that in dynamic cheap talk without feedback, the Sender can achieve a rich set of outcomes by leveraging the Receiver's ability to monitor message frequencies. While the set of achievable payoffs is generally larger than the static partial-commitment benchmark due to detectable temporal correlations, the static benchmark provides a tight characterization for i.i.d. environments and a lower bound for general Markovian environments. Crucially, the dynamic interaction allows the Sender to achieve the full Bayesian persuasion payoff when preferences are state-independent, a result previously thought impossible without feedback.